April 2026 was supposed to ship GPT-6. It shipped GPT-5.5, Claude Opus 4.7, and seven other frontier models instead.

GPT-5.5 ("Spud"): What Actually Shipped on April 23

OpenAI originally framed Spud as GPT-6, with an April 14 launch date. Spud actually landed nine days late on April 23, and OpenAI relabelled it as GPT-5.5. The reading among partners is that the gap over GPT-5.4 (released March 5) was meaningful but not generation-defining, and that the version-number conservatism is consistent with how OpenAI now frames progress around incremental capability gains rather than major leaps. True GPT-6 has no architecture, no parameter count, and no date.

What actually shipped, as of mid-May:

• Three variants in a single rollout. GPT-5.5 (standard), GPT-5.5 Thinking, and GPT-5.5 Pro went live April 23 in ChatGPT and rolled into the API on April 24. Pro is gated to Pro, Business, and Enterprise tiers.
• Pricing doubled. GPT-5.5 Thinking lists at $5 input / $30 output per million tokens, up from GPT-5.4's $2.50 / $15. GPT-5.5 Pro lists at $30 / $180. The Batch and Flex tiers cut 50%; Priority adds 150% on top. The price hike is the sharpest in the GPT-5.x line and ends the year-long downward trend in frontier-model cost per token.
• 1M token context in the API, 400K in Codex. Not the 2M figure that leaked from partner briefings in early April. The 400K Codex cap is the practical ceiling for whole-codebase workflows.
• 88.7% SWE-Bench Verified, 82.7% Terminal-Bench 2.0. Intelligence Index of 59, second only to Grok 5 in published third-party rollups. The headline result is on long context: MRCR v2 at 1M tokens jumps from GPT-5.4's 36.6% to 74.0%, and at 128K to 256K tokens it scores 87.5% versus Claude's 59.2%.
• Super-app remains the strategic story. GPT-5.5 is the engine OpenAI is pointing at the merge of ChatGPT, Codex, and the Atlas browser into a single desktop surface. The 1M context plus the recall numbers make a credible argument for that direction.

Practical read: GPT-5.5 Thinking is the new frontier default if you can absorb the doubled cost and your workload benefits from the long-context jump. For everyday agent work where tool-call latency dominates, GPT-5.4 stays competitive and is now meaningfully cheaper than its successor. Claude Opus 4.6 and GLM-5.1 still hold their ground on coding-heavy benchmarks; the field is no longer two horses.

Claude Opus 4.7: The Public Anthropic Release of the Month

Anthropic shipped Claude Opus 4.7 to general availability on April 16, a direct upgrade to Opus 4.6 at unchanged pricing ($5 in / $25 out per million tokens). Available day-one across Claude.ai, the API, Amazon Bedrock, Vertex AI, and Microsoft Foundry.

• SWE-Bench Verified moved from 80.8% to 87.6%, and SWE-Bench Pro jumped from 53.4% to 64.3%, the largest single-release Pro gain Anthropic has shipped. Four tasks on their internal 93-task coding benchmark that neither Opus 4.6 nor Sonnet 4.6 could solve are now solvable.
• New "xhigh" reasoning effort tier for jobs you would otherwise hand to GPT-5.5 Pro. Latency and token cost both scale, but the agentic loop quality is the headline.
• 3.3x higher-resolution vision. Images up to 2,576 pixels on the long edge (~3.75 megapixels), which makes screenshot-of-an-IDE workflows materially more reliable.
• Updated tokenizer. Same dollar prices, but the same input can map to roughly 1.0-1.35x more tokens than Opus 4.6. Profile real bills before assuming the price is flat.
• 200K context, not 1M. Anthropic did not match GPT-5.5's 1M window here. If you need whole-monorepo passes in a single call, GPT-5.5 Thinking is still the answer.

Read alongside GPT-5.5 a week later, the strategic contrast is clear: OpenAI paid for a long-context leap with a doubled token bill, Anthropic shipped the coding gains at last month's price. For everyday agent work and tool-call-heavy harnesses, Opus 4.7 is the cheaper, steadier pick today.

Claude Mythos: Anthropic's Gated Preview

Anthropic announced Claude Mythos Preview on April 7, available exclusively through Project Glasswing to roughly 50 partner organizations. The focus is on cybersecurity vulnerability detection, reasoning, and coding. Reported numbers include 93.9% on SWE-Bench Verified and 94.6% on GPQA Diamond, both above the publicly available Opus 4.7.

Mythos is described as a step change above Opus 4.7. Preview pricing is steep at $25/$125 per million input/output tokens, reflecting the gated early-access nature of the program; access stays restricted because Anthropic judged broader release too risky after Mythos surfaced thousands of zero-day vulnerabilities in internal evals.

No public release date has been announced. For most developers, Claude Opus 4.7 and Sonnet 4.6 are the current Anthropic options.

Google Gemma 4: Open-Source Gets Serious

Google released the Gemma 4 family on April 2 under Apache 2.0, delivering four models purpose-built for different deployment scenarios:

• Gemma 4 31B Dense the flagship, with benchmark scores that outperform models 20 times its size
• Gemma 4 26B MoE mixture-of-experts variant for efficient inference
• Gemma 4 E4B consumer GPU and edge deployment
• Gemma 4 E2B smartphones and Raspberry Pi devices

All four models support 256K context windows, native vision and audio processing, and fluency in over 140 languages. They are purpose-built for advanced reasoning and agentic workflows.

With over 400 million cumulative Gemma downloads, this release under Apache 2.0 (upgraded from earlier, more restrictive licenses) represents a strategic shift in Google's open model approach.

Zhipu GLM-5.1: China's MIT-Licensed Giant

Zhipu AI released GLM-5.1 under the MIT license, a 744-billion parameter mixture-of-experts model with 40 billion parameters active per forward pass and a 200K context window.

The headline claim: on SWE-Bench Pro, GLM-5.1 reportedly beat both Claude Opus 4.6 and GPT-5.4. Alongside GLM-5.1, Zhipu also released GLM-5V-Turbo, a multimodal variant optimized for coding tasks.

The MIT license makes this one of the most permissive releases of a frontier-scale model to date. No usage restrictions, no registration required.

Alibaba Qwen 3.6-Plus: 1M Context for Agents

Alibaba's Qwen 3.6-Plus targets agentic coding workflows with a 1 million token context window. The model is designed for tasks that require understanding and modifying large codebases in a single pass.

This positions Qwen 3.6-Plus as a direct competitor to Claude Opus 4.6 and GPT-5.4 for the growing market of AI-powered coding agents.

Meta Llama 4: The 10M Token Context Window

Meta's Llama 4 family includes two headline models:

• Llama 4 Scout with a 10 million token context window, the largest of any model released this month
• Llama 4 Maverick with 400 billion parameters, 1 million token context, and native multimodal capabilities

Both models use a mixture-of-experts architecture and are natively multimodal from training (not bolted-on vision after the fact). Meta is using controlled licensing agreements for Llama 4, distinguishing its approach from fully permissive open-source releases.

Arcee Trinity: 400B Under Apache 2.0

Arcee AI released Trinity, a 400 billion parameter model under Apache 2.0. Trinity is designed for enterprise use cases where teams need a large, capable model they can run and modify without licensing restrictions.

Which Model Should You Actually Pick This Month?

With Spud shipped and the open-weight side of the field crowded, which one do I use today is the question that matters more than the headline benchmark numbers. Practical picks by use case, as of mid-May:

Long-context coding (whole-codebase analysis, multi-file refactors). GPT-5.5 Thinking is the new frontier default at 1M tokens; the MRCR v2 jump from 36.6% to 74.0% makes whole-repo passes meaningfully more reliable, but the price doubled. Claude Opus 4.7 caps at 200K context, but it leads SWE-Bench Pro at 64.3% and stayed at last month's $5/$25 pricing - if your refactor fits, it is the cheaper, steadier pick. GLM-5.1 is the value pick if you want frontier-level capability at open-weight cost. Qwen 3.6-Plus is a strong tool-calling alternative with a 1M context and a different token bias.

Local model on a Mac. Gemma 4 31B Dense for any machine with 64GB+ unified memory (M3 Max, M4 Pro/Max). Drop to Gemma 4 26B MoE on 36-48GB. E4B and E2B are the picks for 16-24GB MacBook Airs; both run at usable speeds via mlx and llama.cpp. Llama 4 Scout is the one to load when you need the 10M context for a single document and can absorb the disk footprint.

Agentic browser and desktop work. Tool-call reliability and latency matter more than raw IQ here. Claude Opus 4.7 is the current default for production agents - the 4.7 release specifically improved file-system memory across multi-session agent work, and pricing did not move. GPT-5.5 Thinking helps when a single task spans more than a handful of tool calls or needs the 1M-token window. GLM-5.1 is competitive enough to A/B on your real workflows. The harness matters more than the model. Fazm's custom API endpoint accepts any of these, so you can swap Opus 4.7 for GPT-5.5 Thinking or GLM-5.1 mid-task and watch the failure rate change in place.

Privacy-sensitive workflows (legal, medical, internal HR data). Local Gemma 4 31B Dense or Llama 4 Maverick via Ollama or LM Studio. The cost stops scaling with usage the moment you stop paying per million tokens, which makes large-volume document review and inbox triage economically reasonable for the first time. The GPT-5.5 price hike makes this calculus more obvious, not less.

Frontier reasoning, "I just need the smartest thing in the room." GPT-5.5 Pro at $30/$180 if cost is no object and you want the highest published Intelligence Index outside Grok 5. Claude Opus 4.7 at the new "xhigh" effort level for hard engineering reasoning at a fraction of Pro's cost. Claude Mythos is still Glasswing-partners-only with no public timeline, and the $25/$125 preview pricing is not the everyday-default tier.

ChatGPT-Atlas-style "browse + code + converse" super-app. GPT-5.5 is the engine OpenAI is pointing at this surface. The architecture story is now real, but the migration cost is real too: at $5/$30 you should profile your actual token mix before assuming the doubled price is paid back by the long-context performance.

Running Any Of These Models On Your Mac

If you are reading this list trying to pick which model to actually use for day-to-day work, the blocker is almost never the model. It is the harness. A 1M token context window and 95% HumanEval do not matter if the thing cannot open your browser, read your screen, and click the right button.

Fazm is the macOS-side answer to that harness problem. It is a local computer-use agent that drives your actual Mac through the accessibility APIs (not screenshots), and it is model-agnostic:

• Point it at any of the models in this post. Fazm supports custom API endpoints, so you can route through GPT-5.5 Thinking, Claude Opus 4.7, a local Gemma 4 31B, or a GLM-5.1 instance hosted behind a corporate proxy. Same agent, different backend. Swap mid-task to A/B failure rates on your real workflow.
• It works on the apps you already use. Browser, Google Docs, Sheets, Calendar, your CRM, your invoicing tool, Mail, the Finder. Not a headless Chromium in a data center.
• Voice-first. You describe what you want and it does it. No prompt engineering chat loop.
• Fully open source, runs locally. Source is at github.com/m13v/fazm. Your screen and mic never leave your machine unless you explicitly point it at a hosted model.

It runs on macOS 14 or newer. Native Swift/SwiftUI, no terminal required.

Why the harness matters more than the model

Every new benchmark in the tables above is measured in a sanitized sandbox. The moment you point a model at a messy real desktop, the failure modes change. Latency shifts to accessibility-tree queries and app focus transitions, not token generation. Errors cluster around stale UI state, modal dialogs, and half-loaded pages, not reasoning. A model that wins SWE-Bench Pro can still fail to close a Zoom notification blocking the Send button.

The practical bottleneck for small-business automation right now is whichever agent loop plus accessibility plumbing handles those failure modes most gracefully. That is what Fazm optimizes for. The model is a swappable component.

Looking Ahead From Mid-May: What Is Still Pending

Closed and open items now that the dust has settled:

• GPT-5.5 (Spud) shipped April 23, but as a 5.5 not a 6. True GPT-6 has no architecture paper, no parameter count, no pricing, no date.
• Grok 5 from xAI is still in active training on Colossus 2. Public-beta consensus is late Q2 or Q3 2026; prediction markets give roughly a 33% chance of shipping by June 30. xAI's last on-record update was the January 28 Series E announcement.
• Claude Mythos remains gated to the Project Glasswing cohort with no public availability timeline. Anthropic has expanded the technical write-up at red.anthropic.com but has not committed to a wider release window.
• Claude Opus 4.7 shipped April 16 at the same price as Opus 4.6, with SWE-Bench Verified up to 87.6% and SWE-Bench Pro up to 64.3%. If you were on Opus 4.6, the upgrade is a no-brainer; just budget for the new tokenizer (1.0-1.35x more tokens for the same input).
• The 1M token context window beta for Claude Sonnet 4.5 and Sonnet 4 retired on April 30. Sonnet 4.6 and Opus 4.7 are the migration targets; anything still on the beta should already have moved.

April 2026 is the month frontier pricing turned the other way for the first time in a year, and the month open-weight credibility finally caught up on coding benchmarks via GLM-5.1. The volume of shipped models means developers have more genuinely frontier-tier options at a wider cost spread than at any prior point. The best time to evaluate which one fits your use case is now, while the field is fresh and the pricing pages are still settling.

If you want to evaluate them actually doing things on your Mac (not in a chat window), grab Fazm for free. Plug in whichever of these models you are curious about and point it at a real workflow you would normally do by hand, invoicing, CRM updates, inbox triage, scheduling. That is the honest benchmark.